An AI done list is rewiring how work gets assigned

AI done – Before coffee, the author checks what a set of AI agents completed overnight—security sweeps, software updates, and long analyses once done by hand. The change is real, but the acceleration brings a new bottleneck: humans now get overloaded with AI-generated t

Most mornings now, the first thing is not coffee or social media. It’s a phone check—what got done while he slept.

Last Tuesday. that overnight routine carried him through a 2. 000-word briefing on the risks of helium shortages to Asian semiconductor companies for consulting clients. The report covered a dozen companies in some detail. touched on second-order effects. and. most importantly. it came with a surprise: the nightly security audit of his software platforms found a small bug and fixed it. Another message confirmed his daughter had received her French reading practice. And before dawn, a huge customer segmentation analysis simulated 40 people reviewing a proposed new product.

The order of his days has shifted so much that even the language gives him away. The passive voice in that paragraph was intentional. He didn’t complete any of that work. A small staff of AI agents did.

R Mini Arnold—RMA for short—was set up in February. The “R” is a nod to Isaac Asimov. whose robot names often begin with R. and they serve humans without an agenda. “Arnold” is the T-800 from Terminator 2, not the killer, but the reprogrammed protector: tireless, literal, utterly loyal. “Mini” is a tribute to the popular tiny Mac, an elegant desktop box.

RMA coordinates several other agents working in parallel: one editor. another a researcher/analyst. a third takes an investor’s lens. and a fourth focuses on data. On some evenings, up to 10 general coding agents join them. Together, they handle whatever he queues up before he goes to bed. RMA determines how to deploy the squad for security sweeps. software updates. and the tedious technical housekeeping that used to eat into his attention.

The customer segmentation report wasn’t perfect. He spent 10 minutes tweaking it before sharing it with his managing director. If RMA hadn’t done the work, he said, he wasn’t sure when they would have tackled it. Hiring an agency would have cost many thousands of dollars and taken too long.

Every night, tasks like that run to close “open loops” in his mind—the couldas, wouldas, shouldas that can torment a small-business owner. For him, the payoff isn’t just speed; it’s decision-making. He makes decisions more quickly because the analyses are already there.

After only a few weeks, he says, he felt like he had caught up with things planned months later. That industriousness brought an odd new problem: Silicon Valley investor Tomasz Tunguz calls it the “done list.” And having a done list, he says, feels strange.

What now—when the backlog disappears?

His description of the last two years of AI adoption is blunt. For most people, using AI hasn’t changed much. Colleagues call it “fancy Google”: type something into ChatGPT, get a response, type back. More sophisticated users can ask for a research summary or a spreadsheet. Until very recently, he says, people couldn’t hand anything complex to the AI and walk away.

That changed in November 2025, when Anthropic released Opus 4.5. The new model could reliably follow up to roughly 100 instructions. It was best at writing code, but “pretty good” at other things. Less than three months later, Anthropic released Opus 4.6, increasing the model’s capacity.

The pace isn’t just about models. A research group called METR has been tracking “time horizons”—how long and complex a task an AI agent can execute without human intervention. In early 2024, the time horizon was limited to a few minutes; a year later, about 20 to 30 minutes. Soon, agents should handle a daylong task reliably, and within a few years—or less—that might stretch to weeks.

He compares it to Moore’s law. except instead of tracking transistors. METR is tracking time: “specifically. how much of a knowledge worker’s day an AI can own without supervision.” Moore’s law doubled computing power every two years; the “Time Horizon law. ” he writes. doubles cognitive reach every four months.

And it’s accelerating.

RMA doesn’t run like a one-off chatbot. To be useful. it needs persistent instructions that can extend to several thousand words and include details and data about him—his team’s name and roles at Exponential View. where they research the development of AI in the economy. plus major projects from the startups he’s invested in and his new book. Even personal preferences make their way in: his love for eclectic EDM. These text documents teach RMA how to behave and how to communicate—its “personality,” priorities, and resources.

He assigns high-consequence tasks to RMA—product development. pricing. customers. and his daughter’s revision plans—because if he doesn’t. he won’t care enough to take the time to painstakingly check the agent’s work. To do its job, RMA needs access to the systems he uses: email, calendar, code base, project management, and CRM.

Of course, he says, all of this brings real risks. It can send emails on his behalf, execute code, and browse the web. One tool he uses, OpenClaw, is insecure and really hard to set up. Unleashed on the world in late January. the software prompted Nvidia CEO Jensen Huang to declare that every company needs an OpenClaw strategy. Powerful, he says—but brittle and finicky.

“One Silicon Valley luminary” describes OpenClaw as a Ferrari you have to maintain yourself. RMA breaks at least a couple of times a week. Getting it restarted can take hours. Partly for this reason, RMA lives on its own computer. He can cut the power to it from his iPhone if it goes haywire. though he hasn’t yet had to do that.

Once the work is running at this speed, the morning review changes. It doesn’t involve meticulous verification anymore; there’s simply too much to review. Now he validates: is this roughly in the right direction. did it move their thinking forward. does it open a conversation with a client. Validation requires judgment more than deep attention, and he says his judgment muscle doesn’t tire as quickly.

For his human team, though, it creates a new kind of pressure.

As he powers through his AI-generated to-do list. he’s prompting what he describes as a fire hose of detailed requests for colleagues. The requests aren’t quick emails. They are fully researched documents with trade-offs and implementation plans. The questions arrive with speed: Have they thought about foreign-language editions?. Could they build it?. Wouldn’t it be better if they worked on that?.

The work is no longer just execution. It becomes coordination—how tasks get divided, who adjusts priorities, and who is ultimately responsible for getting the work done.

He says it has him wondering whether he should stop hiding behind the workflow and simply talk to his colleagues.

A survey reflects that same mismatch between readiness and reality. In 2025. McKinsey found that 88% of organizations have adopted AI in at least one function. but only about a third are restructuring workflows around it. The implication he points to is straightforward: the AI is ready, individuals are ready, but the organization itself isn’t. Integrating essential human deliberation into workflows moving at the speed of AI remains a challenge. Right now, humans are the bottleneck.

His examples stretch beyond his own business. In early 2024, Sam Altman predicted a one-person billion-dollar company was imminent, thanks to AI. Matthew Gallagher seems poised to reach the same scale with his two-person GLP-1 telehealth startup, Medvi, in 2026. Every enterprise function of Medvi—platform code, website copy, ad creative, customer service—is accomplished via agentic AI.

Then he asks what happens at a different scale: what does that mean for a 20-person company, or a 200-person company? He says he has a team of humans who work for him, but he also has AI agents doing a team’s work—and it’s costing him a few thousand dollars a month.

In that model, workers need to know how to direct, monitor, and manage AI agents. Enterprises also have to invent new workflows.

The emotional turn comes when he admits he felt guilty—not toward the agents, but about himself. One recent evening, lying in bed, he felt guilty about not having enough work to give the agents. He wasn’t worried the AI couldn’t do enough. The anxiety was that he—someone with “amazing capacities” at his disposal—had run out of useful things to do. The idea sounded absurd, he says, until he experienced it.

This question—how to tackle the human bottleneck and design around it—won’t disappear as autonomy improves. He expects AI agents will be able to work for days and then weeks at a time on the projects he assigns. That likely creates new bottlenecks where decision-makers operate.

Whoever figures out how to delegate at that speed, he says, will do very well. For those who don’t, he’s less certain.

AI agents time horizon Opus 4.5 Opus 4.6 METR helium shortages semiconductor industry OpenClaw workflow restructuring McKinsey survey done list delegation bottleneck Exponential View RMA